Omni-RAG: LLM-Assisted Query Understanding

This project implements the core ideas of Omni-RAG (Dong et al. 2025): a Retrieval-Augmented Generation (RAG) system for real-world, noisy, and multi-intent queries.

---

Overview

• LLM-based Query Understanding: Rewriting and decomposing user queries into sub-queries
• Intent-Aware Retrieval: Vector search over external knowledge for each sub-query (e.g., FineWeb)
• Reranking & Generation: BAAI/bge-base-en-v1.5 reranker and GPT-4.1-nano for final answer

---

Installation

# Clone repo and install dependencies
pip install -r requirements.txt

# Set up database and environment variables
cd docker
docker compose up -d

---

Usage

1. Create a .env file with:
   • TIMESCALE_SERVICE_URL
   • OPENAI_API_KEY
2. Setup DB and upload data:

   python src/upload_vectors.py

3. Run the main pipeline:

   python src/main.py

---

Technologies Used

• Python 3.10+
• TimescaleDB (PostgreSQL)
• Docker
• FineWeb Dataset
• GPT-4.1-nano
• text-embedding-3-small
• BAAI/bge-base-en-v1.5

Python Packages

openai
numpy
timescale-vector
pandas
dotenv
sentence-transformers
transformers
pydantic
datasets

---

📖 Reference

• Dong, G. et al. (2025): Omni-RAG: Leveraging LLM-Assisted Query Understanding for Live Retrieval-Augmented Generation
• Dave Ebbelaar – pgvectorscale‑rag‑solution
• Ran, K. et al. (2025): RMIT-ADM+S at the SIGIR 2025 LiveRAG Challenge

---

Research prototype inspired by the official Omni-RAG pipeline.